Checking what an AI answer says, not just what retrieval found

A relevant passage can still disagree with a claim#

The reranker starts with the retrieval, vector and lexical scores supplied by Laravel. It adds phrase and keyword matches, titles, source type and query intent, then removes duplicates, can cap repeated results from one lesson and returns ranking details. It doesn't load an embedding model.

Verification asks a different question: does the selected material support what the answer actually says? Claim extraction separates the assistant's lead, list items and closing text. For a study pack, it separates the takeaway, must-know points, quiz explanations and correct answers.

The scorer combines keyword overlap with stemmed-token overlap, character trigrams and sentence comparisons. It also checks details such as numbers, dates, names, quoted text, sequence and negation. The component named semantic similarity is built from these heuristics, not a neural entailment model.

Return enough detail for the caller to act#

Thresholds and conflict penalties produce a status, scores and supporting chunk IDs for each unit. The assistant verifier can keep supported units, prune others or report weak support. Study-pack verification also checks that enough required material is supported, including both a quiz explanation and its correct option.

A weak-support response can still contain the original lead text. Laravel therefore reads the outcome, not just the text field: it can substitute review suggestions and checks citations against the support map before constructing the learner response.

Skip modes remain explicit. Tests cover selected ranking and unsupported-answer cases, not a general accuracy rate. Even a correctly matched claim only establishes agreement with the supplied material; it doesn't prove the source itself is true.